Practical tools / Browser only

Agent evaluation
planner.

A small record for a real task. Decide what counts as success, run trials yourself, and keep the costs and corrections in view.

You run the work. This page keeps the record.

It does not connect to a model or execute an agent. These observations describe your recorded trials, not statistical proof or a prediction of future reliability.

01 / Define the trialYour evaluation plan

Task

Use one specific task and fixed inputs.

Model / configuration

Record enough detail to repeat this setup.

Intended output

Describe an output you can inspect.

Permitted actions

Name actions that need permission or must remain out of scope.

Success criteria

Check all criteria before selecting pass. Otherwise use fail or unverified.

Reset method

Include files, conversation state and inputs that must be restored.

02 / Record what happened

Your trial records

Pass and fail mean you checked the output against your criteria. Keep unchecked work unverified. Enter 0 for a measured zero; a blank number is incomplete.

No trials yet. Add your first record after a manual run, or load the labelled synthetic example to explore the planner.

Keep a useful record

Your plan, your choice

Entries stay in this open page until you save or export them. Save only on a device you trust. Loading the saved plan or the example replaces the open plan. Reset clears the open plan.

Copy or download the complete record

CSV includes the plan with each trial. Formula-like text is prefixed with an apostrophe for spreadsheet safety; JSON preserves your original text.

Optional storage in this browser

One saved plan per browser profile and site. Nothing saves automatically. This does not sync or encrypt your entries.

Read alongside your plan

Your trial records

Your plan, your choice

A better question makes a better trial.